Entity focused natural language generation

Abstract

Method includes populating fake value for each of entities, to generate string of fake entity values that correspond to entities; inserting sentinel token between adjacent fake values included in the string to generate first input data; generating, by natural language generation model, natural language sentences based on first input data, natural language sentences including one or more fake values from the string; analyzing natural language sentences to determine whether any fake value from the string is missing; based on the fake value missing, summarizing, using text summarization model, natural language sentences to generate text summary; concatenating the text summary with the fake value, to generate second input data; and generating, by a next sentence generation model, additional natural language sentence using the second input data, the additional natural language sentence including the fake value. Additional natural language sentence is combined with natural language sentences to generate final natural language sentences.

What is claimed is: 
     
         1 . A computer-implemented method comprising:
 populating a fake value for each entity within a set of entities, to generate a string of fake entity values that correspond to the entities, respectively;   inserting a sentinel token between adjacent fake values included in the string of fake entity values to generate first input data;   generating, by a natural language generation model, natural language sentences based on the first input data, wherein the natural language sentences comprise one or more fake values from the string of fake entity values;   analyzing the natural language sentences to determine whether a fake value from the string of fake entity values is missing in the natural language sentences;   in response to determining that the fake value is missing, summarizing, using a text summarization model, the analyzed natural language sentences to generate a text summary;   concatenating the text summary with the fake value, to generate second input data;   generating, by a next sentence generation model, an additional natural language sentence, based on the second input data, wherein the additional natural language sentence comprises the fake value; and   outputting a collection of final natural language sentences formed based on the additional natural language sentence and the natural language sentences.   
     
     
         2 . The computer-implemented method of  claim 1 , further comprising:
 prior to the analyzing:
 tokenizing the natural language sentences into a series of tokens; 
 correcting, using a context correction model, context of the natural language sentences based on the series of tokens to generate corrected natural language sentences; and 
 generating pre-processed natural language sentences based on the corrected natural language sentences, 
   wherein the analyzing is performed on the pre-processed natural language sentences.   
     
     
         3 . The computer-implemented method of  claim 1 , further comprising:
 prior to the analyzing:
 tokenizing the natural language sentences into a series of tokens; 
 correcting, using a context correction model, context of the natural language sentences based on the series of tokens to generate corrected natural language sentences; 
 determining a perplexity score for the corrected natural language sentences; and 
 extracting, from the corrected natural language sentences, natural language sentences having the perplexity score of less than a predetermined perplexity threshold, 
   wherein the analyzing is performed on the extracted natural language sentences.   
     
     
         4 . The computer-implemented method of  claim 1 , further comprising:
 prior to the outputting, combining the additional natural language sentence and the natural language sentences into a text portion;   analyzing the text portion to determine whether another fake value from the string of fake entity values is missing in the text portion;   in response to determining that the other fake value is missing, summarizing, using the text summarization model, the text portion to generate another text summary;   concatenating the other text summary with the other fake value, to generate third input data;   generating, by the next sentence generation model, a second additional natural language sentence based on the third input data, wherein the second additional natural language sentence comprises the other fake value; and   combining the second additional natural language sentence with the text portion, to obtain another text portion,   wherein, in the outputting, the collection of final natural language sentences is formed based on the other text portion.   
     
     
         5 . The computer-implemented method of  claim 1 , further comprising:
 prior to the outputting, combining the additional natural language sentence and the natural language sentences into a text portion comprising a first plurality of natural language sentences;   calculating, using a cross-encoder model, a similarity score between the first plurality of natural language sentences and existing natural language sentences; and   determining the collection of final natural language sentences to include natural language sentences having the similarity score less than or equal to a predetermined similarity threshold, from the first plurality of natural language sentences.   
     
     
         6 . The computer-implemented method of  claim 1 , wherein the outputting comprises at least one from among transmitting the collection of final natural language sentences to an external device and storing the collection of final natural language sentences in a memory. 
     
     
         7 . The computer-implemented method of  claim 1 , wherein the additional natural language sentence and the natural language sentences are combined into a text portion, and
 the computer-implemented method further comprises:
 prior to the outputting, repeating the analyzing and, thereafter, repeating the steps of summarizing, concatenating, generating, and combining, in response to determining that at least one other fake value from the string of fake entity values is missing in the text portion. 
   
     
     
         8 . The computer-implemented method of  claim 1 , wherein the inserting further comprises:
 appending to each fake value a symbol that identifies a type of entity for a corresponding fake value.   
     
     
         9 . A system comprising:
 one or more processors;   a memory that is coupled to the one or more processors and stores one or more instructions that, when executed by the one or more processors, cause the one or more processors to perform a method including:   populating a fake value for each entity within a set of entities, to generate a string of fake entity values that correspond to the entities, respectively;   inserting a sentinel token between adjacent fake values included in the string of fake entity values to generate first input data;   generating, by a natural language generation model, natural language sentences based on the first input data, wherein the natural language sentences comprise one or more fake values from the string of fake entity values;   analyzing the natural language sentences to determine whether a fake value from the string of fake entity values is missing in the natural language sentences;   in response to determining that the fake value is missing, summarizing, using a text summarization model, the analyzed natural language sentences to generate a text summary;   concatenating the text summary with the fake value, to generate second input data;   generating, by a next sentence generation model, an additional natural language sentence, based on the second input data, wherein the additional natural language sentence comprises the fake value; and   outputting a collection of final natural language sentences formed based on the additional natural language sentence and the natural language sentences.   
     
     
         10 . The system of  claim 9 , wherein the method further includes:
 prior to the analyzing:
 tokenizing the natural language sentences into a series of tokens, 
 correcting, using a context correction model, context of the natural language sentences based on the series of tokens to generate corrected natural language sentences, 
 determining a perplexity score for the corrected natural language sentences, and 
 extracting, from the corrected natural language sentences, natural language sentences having the perplexity score of less than a predetermined perplexity threshold, 
   wherein the analyzing is performed on the extracted natural language sentences.   
     
     
         11 . The system of  claim 9 , wherein the method further includes:
 prior to the outputting, combining the additional natural language sentence and the natural language sentences into a text portion;   analyzing the text portion to determine whether another fake value from the string of fake entity values is missing in the text portion;   in response to determining that the other fake value is missing, summarizing, using the text summarization model, the text portion to generate another text summary;   concatenating the other text summary with the other fake value, to generate third input data;   generating, by the next sentence generation model, a second additional natural language sentence based on the third input data, wherein the second additional natural language sentence comprises the other fake value; and   combining the second additional natural language sentence with the text portion, to obtain another text portion,   wherein, in the outputting, the collection of final natural language sentences is formed based on the other text portion.   
     
     
         12 . The system of  claim 9 , wherein the method further includes:
 prior to the outputting, combining the additional natural language sentence and the natural language sentences into a text portion comprising a first plurality of natural language sentences;   calculating, using a cross-encoder model, a similarity score between the first plurality of natural language sentences and existing natural language sentences; and   determining the collection of final natural language sentences to include natural language sentences having the similarity score less than or equal to a predetermined similarity threshold, from the first plurality of natural language sentences.   
     
     
         13 . The system of  claim 9 , wherein the additional natural language sentence and the natural language sentences are combined into a text portion, and
 the method further includes:
 prior to the outputting, repeating the analyzing and, thereafter, repeating the steps of summarizing, concatenating, generating, and combining, in response to determining that at least one other fake value from the string of fake entity values is missing in the text portion. 
   
     
     
         14 . The system of  claim 9 , wherein the inserting further includes:
 appending to each fake value a symbol that identifies a type of entity for a corresponding fake value.   
     
     
         15 . A non-transitory computer-readable memory storing one or more instructions that, when executed by one or more processors, cause the one or more processors to perform a method including:
 populating a fake value for each entity within a set of entities, to generate a string of fake entity values that correspond to the entities, respectively;   inserting a sentinel token between adjacent fake values included in the string of fake entity values to generate first input data;   generating, by a natural language generation model, natural language sentences based on the first input data, wherein the natural language sentences comprise one or more fake values from the string of fake entity values;   analyzing the natural language sentences to determine whether a fake value from the string of fake entity values is missing in the natural language sentences;   in response to determining that the fake value is missing, summarizing, using a text summarization model, the analyzed natural language sentences to generate a text summary;   concatenating the text summary with the fake value, to generate second input data;   generating, by a next sentence generation model, an additional natural language sentence, based on the second input data, wherein the additional natural language sentence comprises the fake value; and   outputting a collection of final natural language sentences formed based on the additional natural language sentence and the natural language sentences.   
     
     
         16 . The non-transitory computer-readable memory of  claim 15 , wherein the method further includes:
 prior to the analyzing:
 tokenizing the natural language sentences into a series of tokens, 
 correcting, using a context correction model, context of the natural language sentences based on the series of tokens to generate corrected natural language sentences, 
 determining a perplexity score for the corrected natural language sentences, and 
 extracting, from the corrected natural language sentences, natural language sentences having the perplexity score of less than a predetermined perplexity threshold, 
   wherein the analyzing is performed on the extracted natural language sentences.   
     
     
         17 . The non-transitory computer-readable memory of  claim 15 , wherein the method further includes:
 prior to the outputting, combining the additional natural language sentence and the natural language sentences into a text portion;   analyzing the text portion to determine whether another fake value from the string of fake entity values is missing in the text portion;   in response to determining that the other fake value is missing, summarizing, using the text summarization model, the text portion to generate another text summary;   concatenating the other text summary with the other fake value, to generate third input data;   generating, by the next sentence generation model, a second additional natural language sentence based on the third input data, wherein the second additional natural language sentence comprises the other fake value; and   combining the second additional natural language sentence with the text portion, to obtain another text portion,   wherein, in the outputting, the collection of final natural language sentences is formed based on the other text portion.   
     
     
         18 . The non-transitory computer-readable memory of  claim 15 , wherein the method further includes:
 prior to the outputting, combining the additional natural language sentence and the natural language sentences into a text portion comprising a first plurality of natural language sentences;   calculating, using a cross-encoder model, a similarity score between the first plurality of natural language sentences and existing natural language sentences; and   determining the collection of final natural language sentences to include natural language sentences having the similarity score less than or equal to a predetermined similarity threshold, from the first plurality of natural language sentences.   
     
     
         19 . The non-transitory computer-readable memory of  claim 15 , wherein the additional natural language sentence and the natural language sentences are combined into a text portion, and
 the method further includes:
 prior to the outputting, repeating the analyzing and, thereafter, repeating the steps of summarizing, concatenating, generating, and combining, in response to determining that at least one other fake value from the string of fake entity values is missing in the text portion. 
   
     
     
         20 . The non-transitory computer-readable memory of  claim 15 , wherein the inserting further includes:
 appending to each fake value a symbol that identifies a type of entity for a corresponding fake value.